At 02/20/2017 08:51 PM, Goldwyn Rodrigues wrote:
On 02/19/2017 11:35 PM, Qu Wenruo wrote:
At 02/20/2017 12:35 PM, Goldwyn Rodrigues wrote:
Hi Qu,
On 02/19/2017 09:07 PM, Qu Wenruo wrote:
At 02/19/2017 09:30 PM, Goldwyn Rodrigues wrote:
From: Goldwyn Rodrigues <rgoldwyn@xxxxxxxx>
We are facing the same problem with EDQUOT which was experienced with
ENOSPC. Not sure if we require a full ticketing system such as ENOSPC,
but
here is a fix. Let me know if it is too big a hammer.
Quotas are reserved during the start of an operation, incrementing
qg->reserved. However, it is written to disk in a commit_transaction
which could take as long as commit_interval. In the meantime there
could be deletions which are not accounted for because deletions are
accounted for only while committed (free_refroot). So, when we get
a EDQUOT flush the data to disk and try again.
IIRC Jeff submitted a qgroup patch to allow unlink to be done even we
already hit EDQUOT.
https://patchwork.kernel.org/patch/9537193/
I think that's a better solution, which is quite like what we do to
handle ENOSPC.
These are two separate issues. This issue is where qg->reserved gets
over-inflated because of repeated deletions and recreations within a
commit_interval.
My fault, that's indeed two different bugs.
And I succeeded to reproduce the bug.
Jeff's patch deals with allowing removal even if the quota is exceeded
so that eventually there can be space freed.
I would suggest you apply Jeff's patch and try to run the script I have
presented.
Thanks,
Qu
I combined the conditions of rfer and excl to reduce code lines, though
the condition looks uglier.
Here is a sample script which shows this issue.
DEVICE=/dev/vdb
MOUNTPOINT=/mnt
TESTVOL=$MOUNTPOINT/tmp
QUOTA=5
PROG=btrfs
DD_BS="4k"
DD_COUNT="256"
RUN_TIMES=5000
mkfs.btrfs -f $DEVICE
mount -o commit=240 $DEVICE $MOUNTPOINT
$PROG subvolume create $TESTVOL
$PROG quota enable $TESTVOL
$PROG qgroup limit ${QUOTA}G $TESTVOL
typeset -i DD_RUN_GOOD
typeset -i QUOTA
function _check_cmd() {
if [[ ${?} > 0 ]]; then
echo -n "$(date) E: Running previous command"
echo ${*}
echo "Without sync"
$PROG qgroup show -pcreFf ${TESTVOL}
echo "With sync"
$PROG qgroup show -pcreFf --sync ${TESTVOL}
exit 1
fi
}
while true; do
DD_RUN_GOOD=$RUN_TIMES
while (( ${DD_RUN_GOOD} != 0 )); do
dd if=/dev/zero of=${TESTVOL}/quotatest${DD_RUN_GOOD} bs=${DD_BS}
count=${DD_COUNT}
_check_cmd "dd if=/dev/zero of=${TESTVOL}/quotatest${DD_RUN_GOOD}
bs=${DD_BS} count=${DD_COUNT}"
DD_RUN_GOOD=(${DD_RUN_GOOD}-1)
done
$PROG qgroup show -pcref $TESTVOL
echo "----------- Cleanup ---------- "
rm $TESTVOL/quotatest*
done
It would be better if we can reduce the reproduce time and submit it as
a fstest test case.
Yes, unfortunately, it is more probabilistic than deterministic. But
yes, reducing the size and increasing the commit_interval can improve
the time.
Signed-off-by: Goldwyn Rodrigues <rgoldwyn@xxxxxxxx>
---
fs/btrfs/qgroup.c | 28 +++++++++++++++++-----------
1 file changed, 17 insertions(+), 11 deletions(-)
diff --git a/fs/btrfs/qgroup.c b/fs/btrfs/qgroup.c
index 4700cac..9ace407 100644
--- a/fs/btrfs/qgroup.c
+++ b/fs/btrfs/qgroup.c
@@ -2093,6 +2093,7 @@ static int qgroup_reserve(struct btrfs_root
*root, u64 num_bytes)
struct btrfs_fs_info *fs_info = root->fs_info;
u64 ref_root = root->root_key.objectid;
int ret = 0;
+ int retried = 0;
struct ulist_node *unode;
struct ulist_iterator uiter;
@@ -2101,7 +2102,7 @@ static int qgroup_reserve(struct btrfs_root
*root, u64 num_bytes)
if (num_bytes == 0)
return 0;
-
+retry:
spin_lock(&fs_info->qgroup_lock);
quota_root = fs_info->quota_root;
if (!quota_root)
@@ -2127,16 +2128,21 @@ static int qgroup_reserve(struct btrfs_root
*root, u64 num_bytes)
qg = u64_to_ptr(unode->aux);
- if ((qg->lim_flags & BTRFS_QGROUP_LIMIT_MAX_RFER) &&
- qg->reserved + (s64)qg->rfer + num_bytes >
- qg->max_rfer) {
- ret = -EDQUOT;
- goto out;
- }
-
- if ((qg->lim_flags & BTRFS_QGROUP_LIMIT_MAX_EXCL) &&
- qg->reserved + (s64)qg->excl + num_bytes >
- qg->max_excl) {
+ if (((qg->lim_flags & BTRFS_QGROUP_LIMIT_MAX_RFER) &&
+ qg->reserved + (s64)qg->rfer + num_bytes >
qg->max_rfer) ||
+ ((qg->lim_flags & BTRFS_QGROUP_LIMIT_MAX_EXCL) &&
+ qg->reserved + (s64)qg->excl + num_bytes >
qg->max_excl)) {
+ if (!retried && qg->reserved > 0) {
Does the qg->reserved > 0 check has any extra meaning here?
qg->reserved greater than zero means the qgroup tree is dirty and needs
to be committed to disk. It may or may not have quota reductions which
can only be confirmed after free_ref_root is called in commit. While it
may be a small probability, it will save I/O processing when the quotas
have been reached and more is demanded again and again.
If qg->reserved equals to 0, IIRC it won't goes here anyway.
Or did I miss something?
In this loop the checks are performed. qg->reserved is incremented in
the next loop once we find all the qgroups to be incremented. Why do you
say the instruction pointer will not be in the area if qg->reserved is
zero. What if the qgroup tree is read from the disk or committed to disk
and the quotas have been met/exceeded?
Right, I forgot that reserve is increased in next loop.
Reviewed-by: Qu Wenruo <quwenruo@xxxxxxxxxxxxxx>
Thanks,
Qu
Thanks,
Qu
+ struct btrfs_trans_handle *trans;
+ spin_unlock(&fs_info->qgroup_lock);
+ btrfs_start_delalloc_roots(fs_info, 0, -1);
+ trans = btrfs_join_transaction(root);
+ if (IS_ERR(trans))
+ return PTR_ERR(trans);
+ btrfs_commit_transaction(trans, root);
+ retried++;
+ goto retry;
+ }
ret = -EDQUOT;
goto out;
}
--
To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
